Aggregate Models, Not Explanations: Improving Feature Importance Estimation
This new method could finally make AI trustworthy for critical biomedical research.
Deep Dive
A new paper tackles a major problem in using AI for science: unstable and unreliable feature importance estimates. The research shows that for complex models, the standard approach of aggregating explanations is wrong. Instead, building an ensemble of models first and then explaining the single ensemble provides significantly more accurate results. This was validated on classical benchmarks and a large-scale UK Biobank proteomic study, proving its utility for real-world scientific applications.
Why It Matters
It makes AI models more reliable and interpretable for high-stakes fields like medicine and biology, enabling true scientific discovery.